Reproduce “When Steering Shifts Discourse: Identifying and Intervening on Discourse-Level Changes due to Activation Steering in LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shivam Adarsh, Hetianyu Huang, Lau Safin, Maria Maistro, Christina Lioma
The authors measure discourse shifts caused by contrastive activation steering and test a vector aimed at correcting a selected shift. Across three open-weight models, 424 of 495 tests showed significant shifts. In cases with extreme judge-score flips, the intervention restored the target score in 78% of cases, while the steering preference’s sign remained unchanged in 92.8% of successful corrections.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.